dependency-management

name	dependency-management
description	Python dependency and environment management for multi-service or monorepo python backends. Use when: (1) adding, upgrading, or removing a Python package, (2) responding to Dependabot or security vulnerability alerts (GHSA/CVE), (3) creating a new service that needs its own requirements files, (4) debugging pip install failures or Docker build issues related to dependencies, (5) reviewing or auditing the dependency tree, (6) running pip-compile. Enforces the pip-compile locked-file workflow and tiered dependency hierarchy.
allowed-tools	Bash, Read, Write

Dependencies

This skill requires Python 3.8+ and standard library only. No external packages needed.

To install this skill's dependencies:

pip-compile ./requirements.in
pip install -r ./requirements.txt

See ./requirements.txt for the dependency lockfile (currently empty — standard library only).

---

Dependency Management

3. One runtime per service. Each isolated service owns its own ./requirements.txt lockfile.

Plugin & Skill Script Architecture (Hub-and-Spoke)

1. DRY in source - Hub-and-Spoke. One canonical script file lives at plugins/<plugin-name>/scripts/. Skills that need it use a file-level symlink in their own scripts/ directory pointing back to the root (ln -s ../../../scripts/foo.py).
2. File-level symlinks only. Never symlink entire directories. The Bridge Installer (bridge_installer.py) only resolves individual file-level symlinks. Directory-level symlinks are silently dropped by binary packaging tools.
3. Self-contained at install. The installer (bridge_installer.py) resolves all symlinks to physical copies when deploying to .agents/. This ensures every skill is independently runnable regardless of the source mono-repo's presence.
4. Windows Compatibility. The plugin_installer.py uses a 3-tier strategy for Windows:
   • Symlink (if Developer Mode is on)
   • Junction (fallback for directory-level logic, though file-level is preferred)
   • Full Copy (ultimate fallback) This resolution ensures the Hub-and-Spoke pattern works cross-platform.
5. No Cross-Plugin Script Execution. A skill should never execute python ../../other-plugin/scripts/foo.py. If a cross-plugin capability is needed, use Agent Skill Delegation: instruct the Agent to invoke the target skill via the conversation layer. In the mono-repo source, cross-plugin file-level symlinks are acceptable for shared logic that the installer will then resolve.

Repository Layout (Example)

src/
├── requirements-core.in          # Tier 1: shared baseline (fastapi, pydantic…)
├── requirements-core.txt         # Lockfile for core
├── services/
│   ├── auth_service/
│   │   ├── requirements.in       # Tier 2: inherits core + auth deps
│   │   └── ./requirements.txt
│   ├── payments_service/
│   │   ├── requirements.in
│   │   └── ./requirements.txt
│   └── database_service/
│       ├── requirements.in
│       └── ./requirements.txt

Tiered Hierarchy

Tier	Scope	File	Examples
1 – Core	Shared by >80% of services	requirements-core.in	fastapi, pydantic, httpx
2 – Specialized	Service-specific heavyweights	<service>/requirements.in	stripe, redis, asyncpg
3 – Dev tools	Never in production containers	requirements-dev.in	pytest, black, ruff

Each service .in file usually begins with -r ../../requirements-core.in to inherit the core dependencies.

Workflow: Adding or Upgrading a Package

1. Declare — Add or update the version constraint in the correct .in file.

   • If the package is needed by most services → requirements-core.in
   • If only one service → that service's .in
   • Security floor pins use >= syntax: cryptography>=46.0.5

2. Lock — Compile the lockfile:

   # Core
   pip-compile src/requirements-core.in \
     --output-file src/requirements-core.txt
   
   # Individual service (example: auth)
   pip-compile src/services/auth_service/requirements.in \
     --output-file src/services/auth_service/./requirements.txt
   

   Because services inherit core via -r, recompiling a service also picks up core changes.

3. Sync — Install locally to verify:

   pip install -r src/services/<service>/./requirements.txt
   

4. Verify — Rebuild the affected Docker/Podman container to confirm stable builds.

5. Commit — Stage and commit both .in and .txt files together.

Workflow: Responding to Dependabot / Security Alerts

1. Identify the affected package and fixed version from the advisory (GHSA/CVE).

2. Determine tier placement:

   • Check if the package is a direct dependency (appears in an .in file).
   • If it only appears in .txt files, it's transitive — pinned by something upstream.

3. For direct dependencies: Bump the version floor in the relevant .in file.

   # SECURITY PATCHES (Mon YYYY)
   package-name>=X.Y.Z
   

4. For transitive dependencies: Add a version floor pin in the appropriate .in file to force the resolver to pull the patched version, even though it's not a direct dependency.

5. Recompile all affected lockfiles. Since services inherit core, a core change means recompiling every service lockfile. Use this compilation order:

   # 1. Core first
   pip-compile src/requirements-core.in \
     --output-file src/requirements-core.txt
   
   # 2. Then each service
   for svc in auth_service payments_service database_service; do
     pip-compile "src/services/${svc}/requirements.in" \
       --output-file "src/services/${svc}/./requirements.txt"
   done
   

6. Verify the patched version appears in all affected .txt files:

   grep -i "package-name" src/requirements-core.txt \
     src/services/*/./requirements.txt
   

7. If no newer version exists (e.g., inherent design risk like pickle deserialization), document the advisory acknowledgement as a comment in the .in file and note mitigations.

Container / Dockerfile Constraints

• Dockerfiles only use COPY ./requirements.txt + RUN pip install -r ./requirements.txt.
• No RUN pip install <pkg> commands. No manual installs.
• Copy ./requirements.txt before source code to preserve Docker layer caching.

Common Pitfalls

• Forgetting to recompile downstream services after a core .in change.
• Pinning == instead of >= for security floors — use >= so pip-compile can resolve freely.
• Adding dev tools to production .in files — keep pytest, ruff, etc. in requirements-dev.in.
• Committing .txt without .in — always commit them as a pair.